\makeatletter
\@ifundefined{HCode}
{\documentclass[screen,CRGEOS,Unicode,biblatex,published]{cedram}
\addbibresource{crgeos20250220.bib}
\newenvironment{noXML}{}{}
\let\citep\parencite
\let\citet\textcite
\def\xcitealp#1#2{\citeauthor{#1}, \citelink{#1}{#2}}
\def\citealt#1#2{\citeauthor{#1}, \citelink{#1}{#2}}
\newcommand*{\citelink}[2]{\hyperlink{cite.\therefsection @#1}{#2}}
\def\defcitealias#1#2{}
\let\citetalias\textcite
\let\citepalias\parencite
\def\thead{\noalign{\relax}\hline}
\def\endthead{\noalign{\relax}\hline}
\def\tbody{\noalign{\relax}}
\def\tabnote#1{\vskip4pt\parbox{.95\linewidth}{#1}}
\def\tsup#1{$^{{#1}}$}
\def\tsub#1{$_{{#1}}$}
\def\ndash{\text{--}}
\def\hyphen{\text{-}}
\def\og{\guillemotleft}
\def\fg{\guillemotright}
\def\0{\phantom{0}}
\makeatletter
\g@addto@macro{\UrlBreaks}{\UrlOrds}
\gappto{\UrlBreaks}{\UrlOrds}
\RequirePackage{etoolbox}
\def\jobid{crgeos20250220}
%\graphicspath{{/tmp/\jobid_figs/web/}}
\graphicspath{{./figures/}}
\def\xmorerows#1#2{#2}
\newcounter{runlevel}
\usepackage{multirow}
\def\nrow#1{\@tempcnta #1\relax%
\advance\@tempcnta by 1\relax%
\xdef\lenrow{\the\@tempcnta}}
\def\morerows#1#2{\nrow{#1}\multirow{\lenrow}{*}{#2}}
\let\MakeYrStrItalic\relax
\def\refinput#1{}
\def\enlabel#1{\hypertarget{#1}{}}
\let\ubreak\break
\csdef{Seqnsplit}{\\}
\def\botline{\\\hline}
\def\back#1{}
\newcommand*{\hyperlinkcite}[1]{\hyper@link{cite}{cite.#1}}
\DOI{10.5802/crgeos.309}
\datereceived{2025-03-10}
\daterevised{2025-07-29}
\dateaccepted{2025-07-30}
\ItHasTeXPublished
}
{
\PassOptionsToPackage{authoryear}{natbib}
\documentclass[crgeos]{article}
\usepackage[T1]{fontenc}
\def\enlabel#1{\label{#1}}
\def\CDRdoi{10.5802/crgeos.309}
\def\citelink#1#2{\citeyear{#1}}
\def\xcitealp#1#2{\citealp{#1}}
\def\xmorerows#1#2{\morerows{#1}{#2}}
\let\ubreak\relax
\makeatletter
\def\hyperlinkcite#1#2{\citeyear{#1}}
}
\makeatother

\usepackage{upgreek}

\dateposted{2026-01-28}
\begin{document}

\begin{noXML}

\CDRsetmeta{articletype}{research-article}

\TopicFR{Sciences du climat}
\TopicEN{Climate Sciences}

\editornote{Article submitted by invitation}
\alteditornote{Article soumis sur invitation}

\title{An outlook on the rapid decline of carbon sequestration and
perspectives for an improved monitoring of French forests}

\alttitle{Une perspective sur le d\'{e}clin rapide du puits de carbone
des for\^{e}ts fran\c{c}aises et l'am\'{e}lioration de leur
suivi}

\author{\firstname{Philippe} \lastname{Ciais}\CDRorcid{0000-0001-8560-4943}\IsCorresp}
\address{Laboratoire des Sciences du Climat et de l'Environnement,
LSCE/IPSL, CEA-CNRS-UVSQ, Universit\'{e} Paris Saclay, 91191 Gif-sur-Yvette, France}
\email[P. Ciais]{philippe.ciais@cea.fr}

\author{\firstname{Chuanlong} \lastname{Zhou}\CDRorcid{0000-0001-8848-8247}}
\addressSameAs{1}{Laboratoire des Sciences du Climat et de l'Environnement,
LSCE/IPSL, CEA-CNRS-UVSQ, Universit\'{e} Paris Saclay, 91191 Gif-sur-Yvette, France}

\author{\firstname{Pascal} \lastname{Schneider}}
\addressSameAs{1}{Laboratoire des Sciences du Climat et de l'Environnement,
LSCE/IPSL, CEA-CNRS-UVSQ, Universit\'{e} Paris Saclay, 91191 Gif-sur-Yvette, France}
\address{Swiss Federal Institute for Forest, Snow and Landscape Research (WSL), Forest Dynamics, Birmensdorf, Switzerland}
\address{ETH Zurich (Swiss Federal Institute of Technology), Institute of Terrestrial Ecosystems, Zurich, Switzerland}

\author{\firstname{Martin} \lastname{Schwartz}\CDRorcid{0000-0003-4038-9068}}
\addressSameAs{1}{Laboratoire des Sciences du Climat et de l'Environnement,
LSCE/IPSL, CEA-CNRS-UVSQ, Universit\'{e} Paris Saclay, 91191 Gif-sur-Yvette, France}

\author{\firstname{Nikola}\nobreakauthor\lastname{Besic}\CDRorcid{0000-0002-2597-4042}}
\address{IGN, GeoData, Laboratoire d'Inventaire Forestier (LIF), 54000 Nancy, France}

\author{\firstname{C\'{e}dric} \lastname{Vega}\CDRorcid{0000-0002-2740-8845}}
\addressSameAs{4}{IGN, GeoData, Laboratoire d'Inventaire Forestier (LIF), 54000 Nancy, France}

\author{\firstname{Jean-Daniel} \lastname{Bontemps}\CDRorcid{0000-0003-3045-2799}}
\addressSameAs{4}{IGN, GeoData, Laboratoire d'Inventaire Forestier (LIF), 54000 Nancy, France}

\keywords{\kwd{Forest}
\kwd{Carbon sink}
\kwd{Climate change}}

\altkeywords{\kwd{For\^{e}t}
\kwd{Puits de carbone}
\kwd{Changement climatique}}

\shortrunauthors

\thanks{EYE CLIMA HE project Grant Agreement No. 101081395, French
German AI4FOREST project (ANR-22-FAI1-0002-01).}

\begin{abstract}
In this study, we present and discuss changes in carbon
storage in French forests from 1990 to 2022, derived from CITEPA
statistics on forest carbon accounting. These statistics are primarily
informed by National Forest Inventory (NFI) data collected from
systematic samples of forest plots across Metropolitan France, as well
as additional sources related to forest removals, soils or wood
products. As NFI is designed to provide statistical estimations of
forest growing stock, gains and losses only at the national or
subnational levels but not to deliver detailed spatial outlooks on
disturbances carbon losses from fires, droughts and insect attacks, we
also outline a prospect for future improvements enabled by remote
sensing and the development of multi-source inventories.

At a national level, a continuing removal of CO\textsubscript{2} from the atmosphere
occurred from 1990 to 2022, as harvest and mortality-induced CO\textsubscript{2} losses remained smaller than CO\textsubscript{2}  removals by forest growth and
the increase in forest area (ca. 80 000 ha per year since 2005 but
insignificant in terms of increased carbon stocks at present). The
CO\textsubscript{2} removal by forests was 49.3 MtCO\tralicstex{\textsubscript{2}\textperiodcentered{yr}\textsuperscript{−1}}{$_{2}{\cdot}$yr$^{-1}$} in 1990,
increased to reach a peak of 74.1 MtCO\tralicstex{\textsubscript{2}\textperiodcentered{yr}\textsuperscript{−1}}{$_{2}{\cdot}$yr$^{-1}$} in 2008 and then
quickly decreased down to 37.8 Mton CO\tralicstex{\textsubscript{2}\textperiodcentered{yr}\textsuperscript{−1}}{$_{2}{\cdot}$yr$^{-1}$} in 2022. The
changes in CO\textsubscript{2} removal by forests can be separated into three
phases. From 1990 to 2013, the CO\textsubscript{2} removal increased alongside the
increasing growth of living trees. A spike in carbon loss was caused by
the passage of the Lothar and Martin extra-tropical cyclones but
forests recovered rapidly within a few years. In contrast, from 2013 to
2017, the CO\textsubscript{2} removal by forests quickly decreased due to
increasing CO\textsubscript{2} losses from harvest and natural mortality and a
trend of decreasing productivity (Hertzog L. R. \mbox{et al.,} \textit{Sci.
Total Environ.} \textbf{967} (2025), article no. 178843), each process
contributing almost equally. After 2017, the sink remained low and
mortality rates stayed larger than during any of the previous years.
The recent period is marked by climate shocks such as summer droughts
and heatwaves in 2015, 2018, 2022, 2023. The full impacts of the
droughts in 2022 and 2023 are not yet covered with full precision,
as some of the sites measured by the national inventory before those
droughts are still pending a second visit. Delayed tree mortality can
also manifest years after a drought has occurred.

At a regional level, contrasted trajectories were identified. Southern
Mediterranean regions where forests have a low harvest rate have also
experienced a lower increase in mortality and a sustained CO\textsubscript{2}
uptake. Despite high harvest intensities, the Landes plantations also
show an increasing CO\textsubscript{2} sink. In contrast, all northern regions and
Corsica have seen a strong decline in their CO\textsubscript{2} removal rates,
except in the Ile-de-France region (larger Paris area), where the
CO\textsubscript{2} sink was constant during the last 30 years, possibly because
many forests are used for recreation and are subjected to low harvest
pressure. Two regions, the Hauts-de-France and Grand Est forests, stand
out as becoming net emitters of CO\textsubscript{2} to the atmosphere. Other
regions where the CO\textsubscript{2} sink declined and is now close to zero are
Normandy, Corsica, and Bourgogne-Franche-Comt\'{e}. A detailed analysis
was conducted to identify where trees are dying in France, the regions
with increased mortality, and which species and tree sizes are most
affected.

We conclude with a perspective on how traditional sample-based
statistical estimation of forest carbon changes, as implemented in
classical NFI approaches, can be complemented by high-resolution
satellite and LiDAR data, together with denser monitoring of mortality
processes. Progress in remote sensing technologies supports both
model-based approaches aimed at mapping the carbon budget and enhanced
inventory techniques for accurate estimation at finer spatial scales.
Given the limited continuity of some long-term forest flux estimates,
we finally outline potential pathways to strengthen carbon sink
quantification in the near future.
\vspace*{-2pt}
\end{abstract}

\begin{altabstract} 
Dans cette \'{e}tude, nous pr\'{e}sentons et analysons
les changements intervenus dans le stockage du carbone des for\^{e}ts
fran\c{c}aises entre 1990 et 2022, \`{a} partir des donn\'{e}es du
CITEPA sur la comptabilit\'{e} carbone des for\^{e}ts. Ces statistiques
s'appuient principalement sur les donn\'{e}es de l'Inventaire forestier
national (IFN) collect\'{e}es \`{a} partir d'\'{e}chantillons
syst\'{e}matiques de parcelles foresti\`{e}res r\'{e}parties sur
l'ensemble du territoire m\'{e}tropolitain fran\c{c}ais, ainsi que sur
des sources suppl\'{e}mentaires li\'{e}es aux pr\'{e}l\`{e}vements
forestiers, aux sols ou aux produits du bois. Comme l'IFN est con\c{c}u
pour fournir des estimations statistiques du volume de bois sur pied,
des gains et des pertes uniquement au niveau national ou r\'{e}gional,
mais pas pour fournir des cartes d\'{e}taill\'{e}es sur les
perturbations et les pertes de carbone dues aux incendies, aux
s\'{e}cheresses et aux attaques d'insectes, nous donnons \'{e}galement
une perspective d'am\'{e}liorations futures rendues possibles par la
t\'{e}l\'{e}d\'{e}tection et le d\'{e}veloppement d'inventaires
multi-sources. Au niveau national, une absorption continue de
CO\textsubscript{2} atmosph\'{e}rique par les for\^{e}ts s'est produite
de 1990 \`{a} 2022, les pertes de carbone induites par les r\'{e}coltes
et la mortalit\'{e} restant inf\'{e}rieures aux gains de carbone par la
croissance foresti\`{e}re et l'augmentation de la superficie
foresti\`{e}re (environ 80 000 ha par an depuis 2005, mais
insignifiante en termes d'augmentation des stocks de carbone \`{a}
l'heure actuelle). La capture nette du CO\textsubscript{2} par les
for\^{e}ts \'{e}tait de 49,3 MtCO\textsubscript{2} par an en 1990, a
augment\'{e} pour atteindre un pic de 74,1 MtCO\textsubscript{2} par an
en 2008, puis a rapidement diminu\'{e} pour atteindre 37,8
MtCO\textsubscript{2} par an en 2022. Les changements dans l'absorption
du carbone par les for\^{e}ts peuvent \^{e}tre divis\'{e}s en trois
phases. De 1990 \`{a} 2013, l'absorption a augment\'{e}
parall\`{e}lement \`{a} la croissance des arbres vivants. Le passage
des temp\^{e}tes Lothar et Martin a provoqu\'{e} une forte augmentation
des pertes de carbone, mais les for\^{e}ts se sont rapidement
r\'{e}tablies. En revanche, de 2013 \`{a} 2017, l'absorption du
CO\textsubscript{2} par les for\^{e}ts a rapidement diminu\'{e} en
raison de l'augmentation des pertes li\'{e}es \`{a} la r\'{e}colte et
\`{a} la mortalit\'{e} naturelle, ainsi que d'une baisse de la
productivit\'{e} ((Hertzog L. R. et al., \textit{Sci. Total Environ.}
\textbf{967} (2025), article no. 178843)), chaque processus contribuant
de mani\`{e}re presque \'{e}gale. Apr\`{e}s 2017, le puits de carbone
est rest\'{e} faible et les taux de mortalit\'{e} sont rest\'{e}s plus
\'{e}lev\'{e}s que pendant toutes les ann\'{e}es pr\'{e}c\'{e}dentes.
La p\'{e}riode r\'{e}cente est marqu\'{e}e par des chocs climatiques
tels que les s\'{e}cheresses estivales et les vagues de chaleur de
2015, 2018, 2022 et 2023. Les impacts complets des s\'{e}cheresses de
2022 et 2023 ne sont pas encore couverts avec pr\'{e}cision, car
certains des sites mesur\'{e}s par l'inventaire national avant ces
s\'{e}cheresses sont toujours en attente d'une deuxi\`{e}me visite. La
mortalit\'{e} diff\'{e}r\'{e}e des arbres peut \'{e}galement se
manifester plusieurs ann\'{e}es apr\`{e}s une s\'{e}cheresse. Au niveau
r\'{e}gional, des trajectoires contrast\'{e}es ont \'{e}t\'{e}
identifi\'{e}es. Les r\'{e}gions du sud o\`{u} les for\^{e}ts ont un
faible taux de r\'{e}colte, ont \'{e}galement connu une augmentation
moindre de la mortalit\'{e} et une absorption soutenue de
CO$_{2}$. Malgr\'{e} une intensit\'{e} de r\'{e}colte
\'{e}lev\'{e}es, les plantations des Landes affichent \'{e}galement un
puits de CO$_{2}$ en augmentation. En revanche, toutes les
r\'{e}gions du Nord et la Corse ont connu une forte baisse de leurs
puits de carbone \`{a} l'exception de l' \^{I}le-de-France o\`{u} il
est rest\'{e} constant au cours des 30 derni\`{e}res ann\'{e}es,
peut-\^{e}tre parce que les for\^{e}ts y sont utilis\'{e}es \`{a} des
fins r\'{e}cr\'{e}atives et soumises \`{a} une faible pression
d'exploitation. Deux r\'{e}gions, les Hauts-de-France et le Grand Est,
se distinguent par leur statut de pertes nettes de carbone. Les autres
r\'{e}gions o\`{u} le puits de carbone a diminu\'{e} et est
d\'{e}sormais proche de z\'{e}ro sont la Normandie, la Corse et la
Bourgogne-Franche-Comt\'{e}. Une analyse d\'{e}taill\'{e}e nous permet
ensuite d'identifier les zones o\`{u} les arbres meurent en France, les
r\'{e}gions o\`{u} la mortalit\'{e} est en hausse, ainsi que les
esp\`{e}ces et les tailles d'arbres les plus touch\'{e}es. Nous
concluons par une perspective sur la mani\`{e}re dont l'estimation
statistique des changements du carbone forestier, bas\'{e}e sur des
\'{e}chantillons, telle que mise en \oe{}uvre dans les approches
classiques de l'IFN, peut \^{e}tre compl\'{e}t\'{e}e par des
donn\'{e}es satellitaires et LiDAR \`{a} haute r\'{e}solution, ainsi
que par une surveillance plus dense des processus de mortalit\'{e}. Les
progr\`{e}s r\'{e}alis\'{e}s dans le domaine des technologies de
t\'{e}l\'{e}d\'{e}tection soutiennent \`{a} la fois les approches
bas\'{e}es sur des mod\`{e}les visant \`{a} cartographier le bilan
carbone et les techniques d'inventaire am\'{e}lior\'{e}es permettant
une estimation pr\'{e}cise \`{a} des \'{e}chelles spatiales plus fines.
Compte tenu de la continuit\'{e} limit\'{e}e de certaines estimations
\`{a} long terme des flux forestiers, nous proposons des pistes de
recherche pour renforcer la quantification des puits de carbone dans un
avenir proche.
\end{altabstract}

%\let\TopAuthors\relax
%\input{CR-pagedemetas}

\maketitle

\twocolumngrid

\end{noXML}

\defcitealias{HCC2025}{ibid.}
\defcitealias{Vallet2023}{ibid.}
\defcitealias{Hlasny2021}{ibid.}

\section{Introduction}\label{sec1}

Forests in France have been absorbing carbon over at least the past 
150~years, mainly due to the delayed effect of favorable tree demography,
following the plantations of new forests and the expansion of trees
over abandoned agricultural lands---particularly in mountain regions,
northwestern plains and southern France
\citep{Lambin2011,DenardouTisserand2019}. This trend reflects a broader
European phenomenon known as the \textit{forest transition}
\citep{Mather1999}. Early afforestation for land protection started in
the 19th century (Landes, Sologne, mountain restoration) and the
long-lasting policy of coppice conversion into standard forests to
deliver timber wood, initiated in 1830, has led to increased forest
density. During this period, harvest pressure has lowered due to fossil
fuel utilization, and remained relatively modest. Natural mortality
rates were low. This transition led to a significantly larger average
stock per hectare, as the demand for fuelwood declined with the
development of alternative energy sources. Consequently, as wood
removal from harvest and wood losses from natural mortality remained
lower than the growth of forests, the result was increased carbon
storage \citep{Bontemps2020}---a trend also reflected in European
forest statistics \citep{Bontemps2021}.

The French national forest inventory aims to provide a reliable
estimation of wood volume and its changes through time by sampling
trees at systematic locations. While it also delivers statistics on
forest carbon stocks \citep{ING2024}, more detailed analyses of carbon
fluxes intended for national and international reporting purposes are
processed by CITEPA (\url{https://www.citepa.org/}), using prior
expertises on the issue \citep[e.g.][]{CARBOFOR2004}. Of note, the NFI
sampling design has changed over time, which currently limits the
set-up of a long-term and harmonized vision of the forest carbon sink.
Between 1961 and 2004, only decennial, departmental, and spatially
asynchronous inventories were implemented, and all forest fluxes were
estimated by retrospective methods. After the severe wind storms
\textit{Lothar} and \textit{Martin} in 1999, extensive systematic and
annual samples of non-permanent plots were established in 2005. The
samples are renewed every year to meet the growing standard of annual
forest inventory, thereby strengthening the forest monitoring frequency
\citep{Bontemps2024}. In 2010, initial samples started being revisited
once after five years (notion of semi-permanent plots, with visits
termed $N$ and $N+5$ in the following) to reliably estimate forest
harvests, in addition to the standard renewal of annual systematic
samples. In 2015, this system was extended to the measurement of
\mbox{mortality} and growth. As a result, flux statistics established
over the long term lack consistency across the full 1990--2022 period
covered in this study. 

This study is based on the publicly available data originating from the
National Forest Inventory and on other data sources, processed by
CITEPA, the national carbon inventory agency. We analyze how carbon
stocks and fluxes causing stock changes in forests have changed between
1990 and 2022, as reflected in these official data. The analysis of the
net carbon sink is complemented by studying the evolution of gains from
forest growth and losses from natural mortality and harvest. The
results are shown at the scale of Metropolitan France and large (NUTS2)
administrative regions. We present and comment on these data, including
net carbon changes and gross gains and losses at national scale in
Section~\ref{sec3}, and on a regional scale in Section~\ref{sec4}. We
show that natural mortality has increased dramatically over recent
years in different regions. We present in Section~\ref{sec5} more
information about where in France trees have been dying, which species
have been dying more frequently, and which size of trees have been more
affected over the last years. Section~\ref{sec6} is dedicated to a prospect on
new mapping and estimation facilities supported by the development of
remote sensing (RS) and machine-learning technologies.

\section{Data sources and methods}\label{sec2}

\subsection{National forest inventory}\label{sec21}

\subsubsection{Sampling design and mensuration protocols}\label{sec211}

The National Forest Inventory (NFI) survey is based on a systematic,
annual, two-phase (aerial and field) and two-stage sampling design,
implemented to provide reliable forest estimates at national and
regional scales in support of public policies. Annual samples are drawn
as gridded subsets from a systematic grid, organized into panels so
that both each individual year and each set of five successive years
maintain systematic spatial coverage across the territory, thereby
ensuring spatial representativeness at multiple temporal scales 
\citep{Bouriaud2023}. 

A typical sampling unit is also called a `forest plot' and has a
maximum radius of 15~m (0.1 ha) for tree assessment and typically
contains between 5 and over 100 trees. During the first visit,
measurements are taken of the diameters of the observable trees
(${\geq}7.5$~cm diameter) at a conventional breast height (1.3~m above the
ground), while the height is measured only for a subsample, with a
statistical model-based imputation which can be applied to the tree
heights that are not measured. Using allometric equations depending on
height and diameter, volumes of standing wood (stem or aerial volume)
in the sampled forest are obtained. Growth is estimated using tree
coring at breast height, focused on the last five years of growth at
the first visit, and is combined with volume allometries to estimate
volume growth. It is supplemented by the total volume of new
recruitments (trees that have grown above a threshold diameter in the
meantime, or \textit{ingrowth} term). Although future growth
information will be derived from differences in tree circumference
measured between successive 5-year inventories, it currently remains
based on tree coring data collected at the initial ($N$) visit,
reflecting growth over the previous five years. \looseness=-1

The trees which were established as cut, dead or windthrown (at the $N+5$
visit) also make part of the ``production'' variable, with the assumption
that they were still growing for 2.5 years before being cut. The field
measurements make an inventory of the trees that died from natural
causes, i.e., were alive during the first visit but were found dead
during the second one. Of note, dead trees that are harvested before
the second visit (as can be the case in e.g.\ massive decline events
such as bark beetle attacks) will not contribute to the mortality, but
rather to the harvest flux. As such, massive mortality can therefore
drive harvest increases, without being entirely identified as such. The
census reports the percentage of trees that were harvested with
evidence of being cut and removed between the first and the second
visit. 

Since 2010, each year, up to 7000 plots have been newly sampled based
on a predefined sampling plan, and up to 7000 plots are re-visited from
the same locations that were sampled for the first time 5 years before,
thus in total, ca.\ 14~000 plots are measured each year. \looseness=-1

{\advance\baselineskip by -.14pt

Importantly, harvest measurements taken before 2010 were exploratory
and long considered unreliable by the NFI, due to inaccuracies in the
stump inventory and dating protocols on temporary plots. Recent
research works  \citep{Audinot2021,DenardouTisserand2019} have
confirmed their negative bias, with historical underestimation by ca.\ 
50\%. For those reasons, harvest rates have been assessed from external
data sources by CITEPA before 2010, including the \textit{Enqu\^{e}te
Annuelle de Branche} from the Agricultural statistics based on wood
fluxes transformed in the sawmills \citep{AGRESTE2022}, and the energy
wood survey of environmental statistics  \citep{LaConsommation2024}.
The temporal continuity of harvest statistics over the study period
(1990--2022) is therefore not guaranteed, and has not been studied so
far.

\vspace*{-2pt}

\subsubsection{Associated statistical inference}\label{sec212}

\vspace*{-1pt}

To obtain the changes in volume and carbon stocks in forests over the
entire country or a large region, a statistical inference method is
applied to the ensemble of the forest plots, with methodologies of
survey sampling \citep{Duong2024} that allow for the computation of an
uncertainty on the estimates. Currently, the precision achieved for the
national estimate of growing stock volume is approximately 1.5\%. On
the scale of smaller regions (e.g., individual forests), the high
heterogeneity of French forests---where 1-ha parcels can host diverse
species, tree sizes, and management histories---results in inevitably
higher uncertainty due to limited sample size. Error quantification can
be accessed in practice using the online tool OCRE
(\url{https://ocre-gp.ign.fr/ocre}).\looseness=-1

Standard forest inventory statistics provide mean changes as five-year
running averages to increase estimate precision. However, this comes
with a known risk of lagging behind evolving forest trends or sudden
events \citep{VanDeusen2011}.
Estimates for successive median years gradually reflect such changes:
for any given disturbance, a growing proportion of the annual samples
includes it over the five years, until its full impact is reported
without bias. For example, if a single-year climate shock causes
significant tree mortality---such as the severe summer drought and
widespread fires in 2022---its full effect will only be fully
captured in the five-year running average by 2027. That said,
statistical inference can also be performed using individual annual
samples for more immediate needs, although this comes at the cost of
greater uncertainty. For instance, the damage assessment following the
Klaus storm in 2009 was based on plot revisits conducted within that
same year \citep{IGN2010}.\looseness=-1

}

In accordance with the NFI resolution, statistically rare events are
prone to more uncertain \mbox{assessments,} as reflected in the variance of
estimation. For instance, in 2022, 60~000 ha of forests were affected
by fire events, corresponding to roughly 0.35\% of the forest area.
With 7000 plots re-measured in 2023, only about 25 plots would fall
within a disturbance domain. This is a major reason why ongoing
research in forest inventory is developing remote sensing-based
inventory, and disturbance-based design approaches as monitoring
options adapted to the current environment \citep{VanDeusen2000}.

\vspace*{-3pt}
\subsection{Citepa}\label{sec22}  

The most recent year for which data are available is 2023, published in
2024. The data are based on the national official report of France to
the UNFCCC compiled by the national inventory agency CITEPA
(\url{https://unfccc.int/documents/627737}). CITEPA provides additional
information on regions of Metropolitan France in the National Inventory
Document (\url{https://unfccc.int/documents/645100}). Data from
different cycles of the National Forest Inventory with changes in
sampling strategies were aggregated by CITEPA. The period 1976 to 2004
has ca.\ 12-year successive inventory cycles with fully renewed samples
performed at the department administrative level (NUTS-3) with a
rotation across departments every year. A non-documented interpolation
method was used by CITEPA to elaborate a regional and nationwide
overview of forest fluxes. The 2005--2009 period was the first period
covered with systematic annually renewed samples. The period 2010 to
current has added a single 5-year revisit with identification of
harvested trees since 2010, and actual tracking of trees that died
since 2015. To deal with those methodology changes the CITEPA
methodology did not use inventories over 1976--2004, but used inventory
from 2005 onwards, provided a point estimate for 1990, and made an
interpolation between 1990 and 2005 (see UNFCCC National Inventory
Document---2024 report \url{https://unfccc.int/documents/645100} page
261). 

\begin{figure*}
\vspace*{2pt}
\includegraphics{fig01}
\vspace*{2pt}
\caption{\label{fig1}(a)~Forest carbon gains and losses and net change in
Metropolitan France. Carbon sequestration or CO$_{2}$ removal from the
atmosphere annually from the growth of living trees in established
forests, the contribution of new forests, the increase in soil carbon
and litter, and storage of wood products are based on either
measurement campaigns from the national inventory updated each year or
models and calculations for soil carbon, or using the life cycle of
products for wood product pools. Carbon gains are reported as negative
numbers in the figure. Carbon losses from harvest (brown), mortality
(red) and large disturbances (storms) are reported as positive numbers.
The net change is the thick black curve from the latest national report
of France, the thin curves are previous estimates published by CITEPA
in previous years. (b)~Comparison between the net carbon sink
in biomass reported by CITEPA (solid line representing a five years
average) and our direct stock change estimate of ${\approx}7000$
forest plots measured each year during each NFI campaign with a five
years running average(dotted line with black dots). The variability in
the dotted curve reflects noise in the statistical sampling to have a
national stock change estimate. Since 2006, the reported emissions
(solid line) correspond to a 5 years smoothing of the direct stock
change (dotted) and the whiskers on the right show the average
uncertainty of annual fluxes (black ${=}$ 1-sigma and grey ${=}$ 2-sigma). 
(c)~Net carbon sink of forests (black) and net greenhouse gas
flux of the whole land use sector including forests and agriculture
(pink) compared to the national neutrality trajectory
(`Strat\'{e}gie Nationale Bas Carbone' SNBC) from SNBC-{2} (pink
lines) revised for the current period and the two future five-years
periods until 2030 and the SNBC-3 (orange lines) with the land use
sector objectives being revised down. The pink curve is a smaller sink
than the forest sink because of agriculture emissions of
CH$_{4}$  and N$_{2}$O. Units are Mton CO$_2$-equivalents.}
\vspace*{2pt}
\end{figure*}

\vspace*{-3pt}

\section{Changes in carbon sinks on a national scale}\label{sec3}
\vspace*{-3pt}
\subsection{Carbon sink from the CITEPA inventory}\label{sec31}

Figure~\ref{fig1}a shows the carbon budget of the forest \mbox{sector} in
France from the national inventory agency CITEPA, with negative values
indicating fluxes of CO$_{2}$ removed from the atmosphere and positive
values indicating fluxes of CO$_{2}$ lost to the atmosphere. A
reporting year $n$ uses inventory campaigns up until year $n-2$ and uses a
five-year moving window for the last five years. The net annual carbon
balance (black curve) demonstrates a net forest carbon sink along the
period. This net carbon sink \mbox{increased} from 49.3~MtCO$_{2}{\cdot}$yr$^{-1}$
in 1990 to a peak of 74.1 MtCO$_{2}{\cdot}$yr$^{-1}$ in 2008. During this
first period, the linear rate of relative increase was 
2.45\%~yr$^{-1}$. After 2008, the sink declined and reached a value of 38.8
MtCO$_{2}{\cdot}$yr$^{-1}$ in 2022. This large decrease in the sink after
2008 is not linear and follows three stages. From 2008 to 2013, a small
decrease in the sink was observed, starting with a drop in 2009 from
the \textit{Klaus} storm. From 2013 to 2017, a rapid and large sink
decrease of 43\% was observed. After 2017, the sink remained stable at
a low value of 37.8 MtCO$_{2}{\cdot}$yr$^{-1}$. Over 2013--2017, we observed
a slight reduction in the growth of established forests (dark green
bars) together with a noticeable increase in tree mortality (red bars).
Since each year, France sends a national communication to report its
emissions and sinks of greenhouse gases to the United Nations
Convention on climate change (UNFCCC), we collected reports published
from 2020 to 2024 from the UNFCCC website to evaluate the consistency
between successive reports. Significant changes are identified, with
the 2021 edition underestimating the recent sink decrease, and the 2023
edition overestimating this decrease compared to the latest 2024
edition (Figure~\ref{fig1}a). These changes reflect the inclusion of new sites
measured by the inventory each year, thus impacting the five years
moving window estimates \citep{VanDeusen2011,Roesch2002}. They are also
sensitive to changes in the protocols or conventions for computing
forest carbon accounting estimates \citep{VanDeusen2011}. Climate
shocks leading to abrupt carbon losses in one year will be smoothed in
time by the 5-years reporting window of inventory. For France, extra
carbon losses from wildfires are however included using a specific
yearly\break approach.


\subsection{Gains}\label{sec32}
\subsubsection{Growth}\label{sec321}

The annual gains were found to be dominated by the growth of
established forests (light green). The growth of those forests has
evolved roughly in parallel with the net carbon gain ($R_{2} =0.74$,
$p<0.05$) from 1990 to 2008, with the increase and decrease in
total growth before and after the maximum of the net carbon sink in
2008, explaining 47.70\% of the variation of the net sink.
Interestingly, the growth of forests reached a maximum in 2014, then
decreased slightly by 5.51\% to reach a minimum in 2017 and remained
stable or slightly increased between 2017 and 2022
\citep[see for a recent detailed analysis of inventory data][]{Hertzog2025}. 
New forests from areas converted from another land use (dark green)
accounted for $9.3 \pm 3.4\%$ of the total annual gain, which is a
CO$_{2}$ removal from the atmosphere of 13.3 MtCO$_{2}{\cdot}$yr$^{-1}$ and
appears larger than what may be expected. The carbon sink of these new
forests steadily declined from 19.49 MtCO$_{2}{\cdot}$yr$^{-1}$ in 1990 to
10.05 MtCO$_{2}{\cdot}$yr$^{-1}$ in 2022, and remained fairly steady until
the years 2000, reflecting a slowdown in the area of new forests each
year during the whole period, to which the disruption of the national
forest fund plantation program (FFN) may have contributed\break to a delay.

\subsubsection{Soils}\label{sec322}
The change in carbon stocks in deadwood, soil and litter is only
reported for new forests created from other land use types (light green
bar in Figure~\ref{fig1}a) using emission factors. For established
forests (dark green bars in Figure~\ref{fig1}a), a neutrality
assumption is made that there is no storage of soil carbon. In
contrast, measurement of soil carbon change at 120 long-term forest
monitoring sites in France (ICP-2 monitoring network, or RENECOFOR---mature 
forests with little management) suggests a large rate of
increase in soil carbon of 1.28 tCO$_{2}{\cdot}$ha$^{-1}{\cdot}$yr$^{-1}$
\citep{Jonard2017}. On the other hand, meta-analysis shows that after a
clearcut and intensive management, up to 20\% of the top soil carbon is
lost to the atmosphere, partly neutralizing the role of a carbon sink
in forest soil. This loss of carbon from management activities is not
included in the National Inventory. 

\subsubsection{Wood products}\label{sec323}
The third component of carbon gain is the storage of wood products in
long-lived pools such as construction materials, furniture and
landfills. This term is not measured but calculated using harvested
\mbox{timber} input data and sectoral models of wood transformation and
lifetime in different pools. Wood products are a carbon sink, which
means that their mass increased over time, but this sink represents a
very small component ($4.9\pm2.60\%$) of the total carbon sink in the
forest sector (blue bar in Figure~\ref{fig1}a) and it decreased from
5.40 Mton CO$_{2}{\cdot}$yr$^{-1}$ in 1990 to 1.09 Mton CO$_{2}{\cdot}$yr$^{-1}$ in
2022. This implies that despite carbon storage in wood products being
recognized as a climate mitigation option, the efficiency of this sink
has decreased in France, probably reflecting an increasing share of
wood harvested being transformed into fuelwood for residential heating
and power production, as well as a lower quality of wood products aimed
at reducing costs, which leads to more wood ending up in landfills. The
limited contribution of forest products to the overall sink is also
reflected in the European Forest strategy, with the priority to favor
the in-situ forest sinks \citep{EuropeanCommission2021}.\looseness=-1

\vspace*{-3pt}
\subsection{Losses}\label{sec33}
\vspace*{-3pt}
\subsubsection{Harvests}\label{sec331}

The annual losses shown as positive numbers in Figure~\ref{fig1}a are
the sum of harvest removals (brown bars) and natural mortality (red
bars) encompassing disturbance-driven mortality (fires, diseases,
insects, drought, frost, small storms) and background mortality induced
by competition between individuals in denser forests, and large
disturbance events (orange bars) mainly from two storms reaching
explosive development rates, Lothar and Martin on December 26--29, 1999
and Klaus on January 23--24, 2009 \citep{Ulbrich2001}. The main factor
of carbon loss remains harvests with a loss of $75.8\pm5.1$ Mton
CO$_{2}{\cdot}$yr$^{-1}$. The removal of wood by harvest has offset about
half of the annual growth during the whole period from 1990 to 2022.
More precisely, during the first period of 1990--2008 when the net
carbon sink increased, the harvest offset $58.5\pm 7.8\%$ of the
growth. During the second period from 2008 to 2017 when the carbon sink
quickly decreased, harvest has offset $52.0\pm 2.7\%$ of the growth.
During the more recent period from 2017 to 2022 with a stable net
carbon sink at a low value, the harvest offset $55.4 \pm 1.4\%$ of the
growth. The stability of this ratio indicates that there has been no
long-term increase in harvest pressure in French forests. Moreover,
during the entire period, variations in harvest explained 26.0\%
($R_{2} = 0.26$) of the variations in the net carbon sink, thus twice
as less as the variations in growth, but together accounting for 75\%
of the changes in the\break net C sink. 

\subsubsection{Disturbance events}\label{sec332}

The effect of salvage harvest consecutive to tree death caused by the
two storms is discernible in Figure~\ref{fig1}a. \textit{Lothar} and
\textit{Martin} in December 1999 laid down 300 million trees
\citep{Abraham2000} that is 7\% of the total growing stock volume,
mainly in the Atlantic region for \textit{Martin}, and in Normandy and
the northern part of France for \textit{Lothar}. This loss was
equivalent to three years of normal harvest \citep{Gardiner2013}. A
fraction of the uprooted and broken dead trees were salvaged and sold
later by the wood industry, showing up as a peak of harvest in 2000 and
2001 in Figure~\ref{fig1}a. A smaller increase of harvest is also
observed after the storm \textit{Klaus} in 2009 which laid down 42
millions of m$^3$ of wood mainly in the \textit{Les Landes} plantation
forest \citep{Pawlik2022}. Mortality from extreme winds during these
cyclones affected large contiguous areas of forests, causing massive
tree losses, while small windblown events during other years may be
under-sampled by the inventory. Analysis of airborne photos, ground
surveys and models made it possible to estimate separately the
immediate carbon losses during the srorms and the legacy carbon losses
from non-harvested dead trees and branches
\citep{IFN2009,LesTempetes2003}. The net CO$_{2}$ emissions (orange
bars in Figure~\ref{fig1}a) represent 5.65\% of the total carbon loss
during the year of the event, with a legacy effect of 5.02\% of the
total carbon loss during the following\break five years.

Besides the two peaks of carbon losses from storms, the second largest
cause of carbon loss after harvest is natural mortality (red bars).
Between 1990 and 2001, natural mortality was stable or increased very
slowly, and represented $13.1 \pm 0.7$ Mton CO$_{2}$ per year,
equivalent to $15.3 \pm 0.9\%$ of the loss due to harvest. After 2013,
mortality accelerated and was multiplied by about a factor of two
between 2013 and 2017. During this period which saw a large decrease of
the net carbon sink, the increase of mortality explained 51.5\% and a
coincident increase of harvest explained 48.5\% of the sink decrease,
respectively. Note that the increased harvest signal could reflect the
salvage harvest of recently dead trees after mortality events.
Intriguingly, the crisis period of dying trees in the French forests
appears to have started before severe droughts and heatwaves recorded
in 2018, 2022, 2023, even though 2015 was marked by a strong water
deficit \citep{Orth2016}. Between 2018 and 2022, despite more frequent
and more severe summer heatwaves and droughts, natural mortality has
remained large but it has not increased further (\mbox{Figure~\ref{fig1}}).
Over this recent period, natural mortality represented 48.7\% of the
losses, an unprecedented loss of wood for the economy. Mortality may
form an even greater fraction of harvest since salvage harvest or
sanitary cuts over insect-affected areas after mortality events will be
classified as harvest by the inventory when cut during the 5 years
preceding the reporting year. An accurate separation of
mortality-driven and management harvests is not possible to date, as it
would require a faster remeasurement of inventory plots than after 5
years, with a substantial impact on the sampling design. The issue is
under consideration at a research level.

Importantly, deadwood has likely increased over time in response to
recent mortality events, but changes in deadwood carbon are not
measured by the inventory, except after the two cyclones and not
reported in national reports. Possibly, deadwood carbon is now
increasing on forest floors, which results in a transient carbon
accumulation in French Forests but will give a legacy carbon emission
in the near future when recent dead woody debris will\break decay.  

\subsection{Carbon sink in biomass national report vs.\ direct stock
change from inventory}\label{sec34}

Figure~\ref{fig1}b compares the net forest carbon sink in biomass of
living trees reported by France with a direct year-to-year carbon stock
change estimation from ${\approx}7000$ forest plots calculated in this
study using forest inventory inference based on annual samples, with a
moving average of the past five years. The variability in our biomass
stock change (dotted curve in Figure~\ref{fig1}b) reflects the
statistical sampling error of the inventory when producing an annual
stock estimate even after applying a 5-year smoothing, and is reflected
by the large error bar on the right-hand side of Figure~\ref{fig1}b (a
reason for which annual estimates are not routinely delivered). We
observe that the \mbox{direct} stock change method gives a smaller \mbox{carbon} sink
smaller than the national report in 2019--2021 but a larger sin in 2022.

\begin{figure*}
\vspace*{-3pt}
\includegraphics{fig02}
\vspace*{-3pt}
\caption{\label{fig2}Same as Figure~\ref{fig1}a but for 13
administrative regions of France indicated in the map.}
\vspace*{-3pt}
\end{figure*}

\subsection{Implications for France's carbon neutrality goals}\label{sec35}

France has adopted a national law on carbon neutrality (Strat\'{e}gie
Nationale Bas Carbone---SNBC) which defines and revises carbon
emissions budgets for different sectors for successive five-year
periods. For land use, land use change and forestry sector (LULUCF),
the first SNBC-1 published in 2015 did not have any specific target.
While other emitting sectors have specific targets, the SNBC-{2} had a
target sink for the LULUCF of ${-}$39 Mton CO$_{2}$e${\cdot}$yr$^{-1}$,
revised to  ${-}$43~Mton CO$_{2}$e${\cdot}$yr$^{-1}$ in 2024 for the
first budget 2019--2023, which is about the magnitude of the forest
sink alone (Figure~\ref{fig1}c) \citep{FrenchGov2020}.
However, this target defined in CO$_{2}$ equivalents also include
emissions and absorptions in the LULUCF sector outside forests, mainly
in croplands and grasslands, which have emitted around 9 Mton CO$_2$e${\cdot}$yr$^{-1}$
in recent years. Hence, the current objective of SNBC-2 for the LULUCF
sector as a whole has not been met, and further goals in 2024--2028 and
2029--2033 shown in red in Figure~\ref{fig1}b are unlikely to be
achieved. The government has reduced ambition in the LULUCF sector and
the SNBC-3 has proposed a more modest sink goal of 
${-}$9 Mton CO$_{2}$e${\cdot}$yr$^{-1}$ for 2024--2028 
and ${-}$18~Mton CO$_{2}$e${\cdot}$yr$^{-1}$ for
2029--2033 \citep{FrenchGov2024}. Will the current forest sink remain
stable, the objective will be met. If it further reduces in view of
increased disturbances, reducing growth \citep{Hertzog2025}, and the
long-delayed effect of forest renewal and afforestation policies
(\url{https://agriculture.gouv.fr/francerelance-le-renouvellement-des-forets-francaises}),
SNBC-3 goals set for the next decade may prove challenging to reach. We
also note that increasing deadwood caused by the increase of natural
tree mortality was recently added in the inventory and resulted a large
upward revision of the carbon sink in French forests, with the increase
in deadwood representing 41\% of the sink instead of 1\% before the
mortality crisis as noted by \citet{HCC2025}. This deadwood carbon will
be decomposed and re-emitted to the atmosphere as CO$_{2}$ in the
coming years, and will thus diminish the carbon sink in the future. In
absence of a detailed description of how deadwood was included by
CITEPA as cited by the \citetalias{HCC2025}, this component was not included
in Figure~\ref{fig1}.  

\section{Changes in carbon fluxes at the regional scale}\label{sec4}


\subsection{Regional carbon fluxes from the CITEPA reports}\label{sec41}


To gain insight into the contribution of each region to the nation-wide
reduction of the carbon sink, we \mbox{analyzed} regional trends.
Figure~\ref{fig2} shows \mbox{carbon} gains and losses for 13 administrative
regions (NUTS2) in France. Carbon removal from the atmosphere in `old
established' forests (light green bar) is inferred from the growth of
living trees; the development of new forests (dark green) is based on a
modeling approach with annual land use change data and emission
factors; the increase in soil carbon and litter remains estimated as at
a national level only for those new forests. Carbon losses are from
harvest, mortality, and large disturbances from storms as in
Figure~\ref{fig1}a. Of note, changes in wood products pools are not
reported per region but they are shown at national scale in
Figure~\ref{fig1}a. 

Contrasted trends were observed across the 13 regions of Metropolitan
France in Figure~\ref{fig2}. The \mbox{Hauts-de-France} and Grand Est forests
have become net emitters of carbon to the atmosphere in recent years,
meaning that they no longer contribute to mitigating climate warming.
Other regions saw a strong decline of their carbon removal rates, which
approached zero like in Normandy, Corsica and
Bourgogne-Franche-Comt\'{e}. There is no straight relationship between
management intensity and the decline of the carbon sink in the west and
the north of France. For instance, the intensively managed plantations
of Les Landes (Figure~\ref{fig2}e), largely reafforested after the
storms of 1999 shows an increasing sink. Also, while quite intensively
managed, broadleaved and conifer forests in Northern and eastern France
show a strong decline (Figure~\ref{fig2}g and 2m). The southern
regions of Occitanie (Figure~\ref{fig2}b) and Provence
(Figure~\ref{fig2}f) where forests are less intensively managed, show a
sustained carbon sink. In contrast, Corsica went from a large sink to
nearly zero, as harvest increased enormously in that region
(Figure~\ref{fig2}k) \citep{Suvanto2025}.\looseness=-1

The net carbon sink reached a peak earlier than the national average
(2008) in Britany (Figure~\ref{fig2}h), Corsica (Figure~\ref{fig2}k),
Centre Val de Loire (Figure~\ref{fig2}d), and Hauts-de-France. The
steepest decline of the sink during the period when the national sink
declined from 2008 to 2017 is of $8.2 \pm 2.0$ Mton CO$_{2}{\cdot}$yr$^{-1}$
($8.7 \pm 9.2\%$ per year). The signature of the two storms is more
apparent at the regional level. Lothar and Martin in 1999 caused the
strongest decreases of the carbon sink in Nouvelle-Aquitaine, Grand Est
and Ile de France followed by a rapid recovery in the four consecutive
years whereas Klaus in 2009 affected mainly Nouvelle-Aquitaine. In each
of these impacted regions, a transient increase of harvest reflecting
salvage wood recovery is observed. In Provence and Corsica which are
the most frequently burned regions \citep{Vallet2023} extreme fire
years of 2009 and 2016--2017 caused a loss of carbon sinks of 28.40\%,
followed by partial recovery. The extreme fires in 2022 in
Nouvelle-Aquitaine with 23163 ha of forest area burned are not yet
fully assessed in the latest 5-year average data available from the
national inventory and will likely show up as a large carbon loss in
that region, estimated by an independent study at 6.23 Mton
CO$_{2}{\cdot}$yr$^{-1}$ \citepalias{Vallet2023}.

Harvest increased dramatically in Corsica after 1998, and in Brittany
and Hauts-de-France, Pays de Loire after 2013, in parallel with a rise
of mortality. It is impossible to assess whether this increase in
harvest is a consequence of sanitary cuts of dead trees or if it is due
to other factors. Interestingly, the ratio of harvest-to-growth is
larger in northern regions (average $0.86 \pm 0.32$) where most forests
are accessible production forests than in southern and mountain regions
(average $0.41 \pm 0.20$) where lack of access and terrain limit the
extraction of wood. The lowest harvest-to-growth ratio is found in
\^{I}le-de-France (0.21), Provence-Alpes-C\^{o}te d'Azur (0.22),
Occitanie (0.31) and the highest one is in the intensive plantation of
Les Landes with a typical harvest rotation of 20 years. 

From Figure~\ref{fig2}, natural mortality increased in all the regions
almost coincidentally around the year 2013, but the magnitude of the
increase differs strongly between regions, with the smallest increase
in Provence-Alpes-C\^{o}te d'Azur (6.46\% per year after 2013) and
Occitanie (7.15\% per year) and the highest in Corsica (10.40\% per
year) and Auvergne-Rh\^{o}ne-Alpes (9.28\% per year). In Grand Est
affected by droughts and massive bark beetle attacks on spruce forests
after 2018, mortality increased dramatically after 2015. In all the
regions, mortality showed a sharp increase between circa 2013 and
2017--2020 and remained stable thereafter, despite severe droughts in
2022 and 2023, but increased carbon losses may be expected from
future inventory campaigns. The reasons why mortality increased in 2013
across a majority of regions is still unclear. Previous heat waves like
the one of summer 2003 did not seem to cause a large increase in
mortality, and the warming rates have been high over the last decades
but did not suddenly accelerate in 2013. It should first be stressed
that mortality encompasses both climate-driven and density-driven
mortality \citep{Charruetal2012}, which are not separated to date. In
view of forest capitalization in France \citep{Bontemps2020},
density-dependent mortality is prone to increase, regardless of climate
change, making direct interpretation uneasy. Second, harvests are being
estimated since 2010 using semi-permanent plots, while they were
estimated from external and indirect sources before from the
Enqu\^{e}te Annuelle de Branche, et enqu\^{e}te bois \'{e}nergie.
However, the years 2010--2015 are not impacted by changes in the
mortality estimation protocols which became implemented in 2015, and
used officially only as of 2019 based on semi-permanent plots. Changes
in the NFI protocol have been implemented to increase the accuracy of
forest fluxes, in a monitoring perspective requested by public
policies, in view of ongoing climate change
\citep{Herve2014}. A
major consequence however lies in the difficulty of obtaining a robust
long-term temporal view of forest changes concomitant to accelerated
climate warming in France, urging an unprecedented effort to homogenize
past data and establishing a firm retrospective reconstitution of
forest carbon\break in France.  

\subsection{Changes in tree mortality across regions, species, and height
classes}\label{sec42}

More insight into tree mortality in France across species, and tree
size classes is presented in this section by analyzing the NFI data
collected during first visits between 2010 and 2018, with re-visits
occurring five years later from 2015 to 2022. Following a revision of
the sampling protocol in 2015, each tree in the sample (619~496 trees
from 50~012 forest plots) was individually monitored to determine
survival or mortality over the five-year re-visit period. This approach
improved the accuracy of tree mortality estimations.

We quantified tree mortality in each plot by using two metrics:
volume-based mortality, measured as the cubic meters of wood lost per
year and hectare forested area (m$^{3}{\cdot}$yr$^{-1}{\cdot}$ha$^{-1}$),
and stem mortality rate, expressed as the percentage of censusable 
stems lost per year (\%-stems yr$^{-1}$). Mortality ($M$)
was calculated using the formula \citep{Kohyama2018}: $M = [1 - [N_{t0}
/N_{t1}]^{[1/t]}] \times 100$, where $N_{t1}$ represents the number of
trees alive at the first visit,  $N_{t0}$ is the number of individuals
that survived between visits, and $t$ is the time interval between
visits (five years). To focus on natural mortality, harvested trees
were classified as survivors, ensuring that the estimated rates
reflected mortality independent of logging. However, it is important to
note that large disturbances, such as storms or bark beetle outbreaks,
are often followed by rapid salvage logging. Due to the five-year
revisit interval, the NFI methodology does not capture these short-term
responses, meaning the reported mortality rates likely underestimate
actual natural \mbox{mortality.} To assess uncertainty, we applied a
bootstrapping approach. For each year and group (e.g., region, height
class, species), we resampled the dataset and calculated the mortality
rate across multiple iterations. The final estimate represents the mean
mortality rate across all bootstraps.

Figure~\ref{fig3} presents the spatial distribution of volume-based
mortality rates due to both biotic and abiotic causes, mapped using a
hexagonal grid. The data revealed a clear increase in tree mortality
across France, particularly after 2018, with the most pronounced
effects in the northeastern regions. When aggregated over larger
eco-regions from GRECO (Grandes R\'{e}gions ECOlogiques)
\citep{IGN2024}
Figure~\ref{fig4}a, all regions show a similar mortality rate from 2015
to 2017 of 0.25 to 0.5~m$^{3}{\cdot}$yr$^{-1}{\cdot}$ha$^{-1}$. In 2018,
mortality rates surged in the Jura (10-fold increase from 2015 to
2023), Vosges (6-fold), and Grand Est (4-fold), in likely response to
the extremely hot drought \citep{Schuldt2020}. In
mountainous regions such as the Alps, Massif Central, and
Pyr\'{e}n\'{e}es, mortality rose less drastically but continuously over
time, reaching rates of approximately 
0.75 m$^{3}{\cdot}$yr$^{-1}{\cdot}$ha$^{-1}$.
Mortality rates along the Mediterranean and Atlantic
coasts are generally smaller than in other regions 
(${<}$0.5~m$^{3}{\cdot}$yr$^{-1}{\cdot}$ha$^{-1}$) 
but still showed a 2- to 3-fold larger
rate in 2023 compared to 2015.

\begin{figure*}
\vspace*{-4pt}
\includegraphics{fig03}
\vspace*{-4pt}
\caption{\label{fig3}Spatial distribution of volume-based tree
mortality rates across France from 2015 to 2023. Mortality rates
increased substantially after 2018, particularly in northeastern
regions, reflecting climate-induced stressors and pathogen outbreaks.}
\vspace*{-4pt}
\end{figure*}

Figure~\ref{fig4}b shows species-specific mortality trends, revealing
that Norway spruce, European ash, and silver fir accounted for the
highest volume-based mortality, with a sharp increase after 2018,
reaching levels up to 40 m$^{3}{\cdot}$ha$^{-1}{\cdot}$yr$^{-1}$. The rise in mortality
among Norway spruce and silver fir is closely linked to both direct and
indirect effects of climate change. A 2019 study on the Vosges forests
attributed increasing mortality rates in both species to reduced water
availability and severe drought events with bark beetles
\citep{Piedallu2022}. Norway spruce, is highly susceptible to drought
stress \citep{Arend2021}, which weakens trees and makes them more
vulnerable to bark beetle infestations \citep{Hlasny2021}. These
outbreaks have intensified in recent years, as climate change has
created more favorable conditions for beetle populations---warmer
winters improve their survival rates, while hotter summers accelerate
their life cycle, allowing additional generations to develop
\citepalias{Hlasny2021}. The dramatic rise in {European} ash
mortality (a 16-fold increase from 2015 to 2023) is primarily driven by
the spread of the {fungal} pathogen {\textit{Hymenoscyphus
fraxineus},} which can cause mortality in up to 85\% of infected ash
trees \citep{Carroll2024}. Unlike defoliating insects, which weaken but
rarely kill trees, bark beetles bore through the bark of their hosts,
disrupting the flow of sap and causing tree death. While beetles have
long been endemic to Europe, they have not been a major source of tree
mortality until recently, when summer droughts have weakened trees'
natural defenses and favored the reproductive cycle of the beetles
\citep{Lange2006}. The resulting outbreaks have had widespread
consequences, affecting not only France but also Belgium, Germany,
Austria, and the Czech {Republic} \citep{Hlasny2021}. 
\looseness=-1

\begin{figure*}
\includegraphics{fig04}
\caption{\label{fig4}Trends in tree mortality across different
categories. {(a)} Volume-based mortality rates trends by
eco-region, showing a sharp rise in the Jura, Vosges, and Grand Est
regions. {(b)} Volume-based mortality trends across species,
with Norway spruce, European ash, and silver fir experiencing the
highest volume losses. (c--d) Mortality trends by tree height
class, illustrating the differential impacts on large versus small
trees (taller trees contributing more to carbon losses despite
lower\%-based mortality rate).}
\end{figure*}

Moreover, mortality rates have risen significantly across other major
tree species in France, including European beech, oaks (pedunculate,
sessile, and downy oak), common hornbeam, and sweet chestnut. Chestnut
is known to suffer from pathogen attacks \citep{Jung2018} and beech
suffers from extreme droughts \citep{Leuschner2020}. From 2015 to 2023,
mortality in these species has at least doubled, and in some cases,
tripled. When looking at stem mortality rate-, ash, chestnut, and
spruce had substantially larger rates than other species. This trend
underscored the growing pressure that \mbox{Europe's} forests face under
climate change \citep{Senf2020} emphasizing the urgent need for
adaptive management strategies and targeted policy measures to
safeguard the critical ecosystem functions these forests provide.

While volume-based mortality rates are particularly relevant for
understanding carbon sink dynamics, they remain biased toward the
mortality of large trees and provide little insight into the survival
of younger tree generations, a stage at which evolutionary processes
are at play through natural selection. Figures~\ref{fig4}c 
and~\ref{fig4}d illustrate trends in both volume-based mortality and
stem mortality rates. While small trees (under 10~m in height) exhibited
lower volume-based mortality rates compared to taller trees (over 25~m),
they have significantly higher stem-based mortality rates. This
reflects natural forest dynamics, where smaller trees die more due to
competition for light rather than external disturbances
\citep{Westoby1984}. The stem mortality rate indicates that mortality
has increased across all height classes, highlighting the broad-scale
impact of recent environmental stressors. Given the current species
mix, this could limit the number of mature trees available as well as
the carbon sink in the coming decades. However, shifts in species
composition, with the establishment and spread of more resilient or
better-adapted species, may influence long-term forest dynamics and
carbon sequestration potential \citep{Wessely2024}.

Although large trees have lower mortality rates compared to smaller
trees, volume-based \mbox{mortality} highlights their outsized contribution to
carbon stock changes. This aligns with the impacts of storms discussed
earlier, as larger trees are more vulnerable to windthrow
\citep{Seidl2017}. Additionally, large trees are highly susceptible to
drought stress due to their expansive canopies, which create a high
evaporative demand and generate strong pressure gradients from soil to
atmosphere \citep{Fensham2019,Bennett2015}. As structural integrity
declines, tall trees become more prone to windthrow and secondary
disturbances. Since tree height is also a factor of vulnerability to
water stress \citep{Koch2004} mortality rates should increase with tree
height under the climatic drought pressure of the past years. In view
of the relationship found in Figure~\ref{fig4}d, density-dependent
mortality in young forests is therefore interpreted to dominate this
response. Further research is here needed to elucidate this response,
which may stem from insufficient cover of larger height classes (beyond
30 to 40~m). \mbox{Filtering} out the {potential} increase in
\mbox{density-dependent} \mbox{mortality} resulting from forest 
densification \citep{DenardouTisserand2019}
climate impacts is also an urgent issue \citep{Taccoen2019}. 

\vspace*{-3pt}

\section{New tools to monitor forests using remote sensing and ground
observations: a prospect}\label{sec5}
\vspace*{-3pt}

\subsection{Limits of traditional forest monitoring}\label{sec51}

Ground-based observations, such as the French National Forest Inventory
surveys, provide high-quality data for monitoring short-(5-yr) to
medium-(10--20~years) term forest dynamics on a national or regional
scale (Figures~\ref{fig1},~\ref{fig2}), for the reason that they have
been designed to support the valuation and formulation of national
forest policies. However, these data have several limitations to the
higher spatio-temporal resolutions that form new challenges for future
forest monitoring \citep{Bontemps2022}. \looseness=-1

First, the temporal resolution for forest flux measurements is five
years, as dictated by the semi-permanent plot protocol, a trade-off
between detecting substantial changes and increasing the resolution
with respect to previous practices (around 12 years in the former
approach). Also, samples are renewed every year, allowing for quick
detection of changes in, at least, state variables, but at the cost of
statistical precision, leading to the standard practice of the
moving-window average in annual inventory \citep{Bontemps2024}. This
low-frequency sampling of fluxes hinders the timely detection of sudden
events affecting e.g.\ the mortality of French forests, such as the
extensive wildfires of 2022, or tree mortality caused by storms and
pest outbreaks, part of it \mbox{being} transferred to harvest fluxes when
occurring earlier than 5 years. A perspective here may be to panelize
the revisits across the 5 successive years, which can be operated at a
constant sampling effort, yet complexifying the design. The practice is
routine in some public surveys, as soon as residence times need be
estimated. 

Second, NFIs have been designed to be statistically representative of
large territories, and do not monitor forests at a fine resolution.
While the associated uncertainty is duly quantified in the variance of
estimation, bias also turns out to be possible if forest events are
disseminated across space to the point where they can be qualified as
``rare'' events, locally. Additionally, higher-resolution is of
increasingly crucial interest in France, where forest parcels are often
small and managed by diverse owners with distinct practices. This
aspect is at the origin of two innovation strategies: (i) model-based
mapping and monitoring facilities, based on remote-sensing products,
and often calibrated with forest inventory data, (ii) enhanced
design-based forest inventory development 
\citep[also termed ``multi-source'' inventory;][]{Tomppo2008}
that preserves the statistical inference capacity associated with the
sampling design, and whereby remote sensing products play an auxiliary
role for increasing estimation precision, providing estimation on much
smaller domains, and developing mapping facilities \citep{Vega2021}.

\subsection{Advances in remote sensing}
\subsubsection{New sensors}

In recent years, remote sensing (RS) technology has undergone
significant advances, particularly with the development of new orbital
sensors that provide data of relevance for forest {monitoring}. Among
them, the Sentinel-2 (S2) mission, part of the European Space
Agency's (ESA) Copernicus Earth Observation Program, captures
multispectral imagery at a 10-meter spatial resolution with a revisit
interval of approximately 6~days in France, making it a valuable tool
for detecting changes in vegetation and canopy structure. Similarly,
Sentinel-1 (S1), another component of the Copernicus program, provides
synthetic aperture radar (SAR) imagery at 10-meter resolution,
\mbox{operating} independently of weather {conditions} and daylight, thus
enhancing forest monitoring capabilities. \looseness=1

Beyond optical and radar remote sensing, LiDAR technology is considered
as one of the most efficient data sources for forest structure
analysis. LiDAR sensors measure the three-dimensional structure of
forests by emitting infrared laser pulses and recording their
reflection from different canopy layers. These measurements provide
critical insights into forest height, biomass, and structural
complexity \citep{Balestra2024}. Since 2018, NASA's Global Ecosystem
Dynamics Investigation (GEDI) mission \citep{Dubayah2020} has delivered
sparse but highly accurate data on forest vertical structures across
the globe, enabling unprecedented assessments of forest height and
canopy density. Specifically, canopy height, one of the most
straightforward metrics derived from LiDAR measurements, has shown
strong correlations with key ecological indicators such as biomass,
biodiversity, and forest health \citep{Dubayah2020,Torresani2023}. 
Furthermore, airborne LiDAR missions, such as those conducted by
national forest agencies, complement spaceborne LiDAR data by providing
denser coverage \citep{Coops2021} at finer scales like the French HD
LiDAR program initiated by the IGN \citep{Melun2024}. Enforcing a
regular HD LiDAR cover over the territory however remains largely
uncertain\break to date.

\begin{figure*}
\includegraphics{fig05}
\caption{\label{fig5}(a) Tree height map of France at 10 m resolution
for the year 2020. (b) Examples at three different locations of height
prediction (left) with the corresponding Google map images from 2020,
2018, and 2019 (right). Brighter colors indicate higher heights.
Figure from \citet{Schwartz2023a}.}
\vspace*{-1pt}
\end{figure*}

Over the past decade, advances in artificial intelligence (AI) and
machine learning, particularly deep learning models \citep{LeCun2015},
have transformed satellite remote sensing research. These frameworks
can efficiently process large-scale datasets and are particularly
effective with unstructured data such as images or sound. This makes
them perfectly adapted for the fusion of multiple remote sensing
datasets, such as GEDI LiDAR data with Sentinel-1 and Sentinel-2
imagery, to generate continuous, high-resolution maps of forest
structure and biomass. The past 5 years have seen an increasing number
of studies using these tools to derive such maps globally
\citep{Lang2023,Pauls2024,Tolan2024}, on the continental scale
\citep{Liu2023}, or national \citep{Fayad2024,Liu2023,Schwartz2023a}
and the regional scale \citep{Favrichon2025,Schwartz2024}. 

\subsubsection{Height and biomass mapping}\label{sec522}

In France, \citet{Schwartz2023a} used GEDI, Sentinel-1, and
Sentinel-2 data with a U-Net model \citep{Ronneberger2015}, a deep
learning approach, to generate a 10 m resolution forest height map for
2020, covering the entire metropolitan territory (Figure~\ref{fig5}a).
This map demonstrated high accuracy, with a mean absolute error of
2.94~m. It enables a detailed understanding of French forest structure
at the stand level, as visible in Figure~\ref{fig5}b. Furthermore,
leveraging allometric equations from inventory data, the authors
produced wood volume and biomass maps at 30 m resolution, offering a
{valuable} snapshot of the carbon stored in French forests in 2020.
\looseness=-1

Satellite and AI-based methods for generating forest height and biomass
maps have shown relatively high accuracy, particularly in temperate
regions like France. However, significant uncertainties remain,
especially in biomass estimation, causing a gap between the initial
resolution of RS signals, and their processing into forest state
variables. Biomass maps rely on allometric equations fitted by broad
forest categories or biomes like in \citet{Schwartz2023a} where two
equations were used depending on the leaf type of the trees
(broadleaf/needleleaf). However, within these broad categories, biomass
is influenced by various factors beyond height, including climate
conditions, tree species, forest management practices, and tree cover.
In addition, volume could be more accurately reconstructed than biomass
from remote sensing, given the large variations in wood density across
species and forest types. These complexities make satellite-based
biomass maps less reliable than the height maps they derive from.
Future models should {incorporate} these additional variables into
biomass predictions or try to map biomass from direct measurements
rather than using height as a proxy. In between, such maps could be
used in model-assisted small area estimation approaches to provide
design-unbiased estimates of forest attributes \citep{Zhang2022}.  
\looseness=-1

\subsubsection{Mapping components of the carbon budget}\label{sec523}

Accurately mapping forest biomass is one of many tasks that can be
addressed using Earth Observation (EO) data combined with deep learning
algorithms. The rise of very high-resolution satellite
imagery---such as Maxar satellites that go up to 15 cm
resolution, SPOT satellites (1.5 m resolution), and the Planet Labs
constellation, which provides 3.5 m resolution images of the entire
globe daily---offers new opportunities to analyze forest
structure changes at the tree level. Some of these private data are
made available for research purposes, such as the annual SPOT mosaics
in France that have been used to produce the open-canopy maps from
\citep{Fogel2024}. These
annual 1.5 m resolution maps have proven high accuracy in detecting
individual tree removal in France's forests. They could significantly
enhance our understanding of forest disturbances and improve the
existing attribution of forest losses
\citep{Fogel2024,VianaSoto2024}. By analyzing disturbance patterns,
they could help differentiate between natural forest losses---such as
windthrows, wildfires, and pest outbreaks---and human-driven factors
like selective logging, clear-cuts, and salvage logging, and further
estimate impacted resources \citep{Sagar2025}.

Forest carbon budget monitoring, meaning an accurate tracking of
biomass losses and gains, is crucial to follow climate-related policies
and reduce greenhouse gases emissions. With their frequent updates,
satellite-based solutions enable this monitoring, and several projects
are already operational, including forest cover loss detection
\citep{Hansen2013}, tropical forest degradation and deforestation
tracking \citep[RADD alerts from][]{Hansen2013,Reiche2021}; TMF
dataset from \citep{Vancutsem2021}, and clearcut monitoring in France
through the SUFOSAT project \citep{Mermoz2024}. Following the pioneer
work of \citet{VegaSt-Onge2008,VegaSt-Onge2009}
the increasing precision of new data and models will allow detecting
growth signals  \citep{VegaSt-Onge2008,VegaSt-Onge2009} as well as
forest losses to estimate forest carbon uptake
\citep{Renaud2017} at national to continental scales. However, few
models have successfully generated consistent height time series
validated with external data over large areas. In France, building on
the study presented in Figure~\ref{fig5}, the same authors attempted to
address this challenge by developing a framework for predicting height
annually in a consistent manner. Figure~\ref{fig6} provides an initial
look at these time-series, where forest growth is visible year to year.
In maritime pine plantations (Figure~\ref{fig6}a) of the Landes forest,
forest parcels clearly show growth between 2018 and 2024, while
clearcuts, represented in red, are also distinct. In contrast, for
mature forests with more complex growth dynamics, changes are less
visible, and biomass accumulation in tree woody mass may be overlooked,
particularly for deciduous oaks (Figure~\ref{fig6}d).

\begin{figure*}
\vspace*{-2pt}
\includegraphics{fig06}
\vspace*{-2pt}
\caption{\label{fig6}Forest height in 2018, 2021, and 2024; Google 
Maps~{\textcopyright}~images; and the difference between 2024 and 2018.
The last column shows the height profiles drawn on the maps. The four
lines correspond to different tree species: (a) Maritime pines (44.54
N, ${-}$1.03 E), (b) Poplar plantation (49.72 N, 3.86 E), (c)~Larches
(44,98 N, 6.47 E), (d) Deciduous oaks (48.30 N, ${-}$3.61 E). Adapted from
\citet{Schwartz2023}.}
\vspace*{-3pt}
\end{figure*}

Another key but more exploratory application is the elaboration of tree
species maps \citep{Beloiu2023}, which could refine height-biomass
allometries and improve our understanding of forest diversity within a
region. By analyzing temporal information from satellite time series,
recent studies have successfully identified tree phenology and linked
it with specific species, as demonstrated in Belgium by
\citep{Bolyn2022}. With large labeled datasets like the Pure Forest
dataset \citep{Gaydon2024}, future research can fully exploit remote
sensing data using advanced deep learning models such as Vision
Transformers \citep{Dosovitskiy2020}. These high-precision tree species
maps would significantly enhance forest management, monitoring, and
carbon stock estimation.

\subsection{A persisting need for field measurements}\label{sec53}

While advances in remote sensing and machine learning are transforming
forest monitoring, ground-truthing remains essential for accurate
calibration and assessment. For example, growth measurements, crucial
for estimating carbon stocks, are only possible in open canopies,
requiring in situ circumference tracking as forests mature
\citep{Wernick2021}. Another promising approach to upscaling mortality
prediction is linking ICP (International Co-operative Programme on
Assessment and Monitoring of Air Pollution Effects on Forests) crown
defoliation data. A recent study using defoliation data from Scots
pines in Switzerland shows that ground-assessed defoliation rates serve
as early-warning signals for mortality \citep{Hunziker2022}. These
insights could be expanded using remote sensing and machine learning,
integrating defoliation monitoring with high-resolution satellite data
like \mbox{Sentinel-2} and GEDI LiDAR \citep{Sagar2025}. Similarly,
Sentinel-2 can be used for early detection of bark beetle
outbreaks but still needs systematic ground-truthing 
\citep{Barta2021}. Additionally, citizen science can play a critical
role, as initiatives such as \textit{Sant\'{e} des For\^{e}ts} already
mobilize volunteers to report pest outbreaks, creating a valuable
database for linking mortality to biotic disturbances and climatic
stress. Public records, analyzed with \mbox{large-scale} \mbox{language} \mbox{models},
could further help reconstruct past disturbance events. 

While accurate geolocation of mortality events is essential for
attribution 
\citep[e.g., the French NFI openly provides coordinates within a
700~m radius of the true location, a prevalent practice across world
NFIs][]{Schadauer2024},
with persisting debates about making such data openly available
\citep{Gessler2024}. For those
NFI placed under the authority of public statistics like in France,
sensitive data access for public research is now permitted and managed
according to the law on statistical secrecy. Another---perhaps more
crucial---issue is that the vast \mbox{majority} of NFI protocols are based on
nested plot designs for the sake of cost reduction (not all the trees
\mbox{being} measured), not primarily intended for remote-sensing calibration,
in contrast with experimental forest plots 
\citep[e.g.][]{Seynave2018}. This
limitation is traded off with the statistical robustness of the
systematic sampling design used in the French National Forest
Inventory, which ensures unbiased estimates at large scales, making it
a cornerstone of forest observation. 

\subsection{Advances in enhanced \textit{multisource} forest
inventory}\label{sec54}

While field observation can be used for the calibration and validation
of RS products, it can also \mbox{benefit} from these to satisfy both
increased precision on small domains, and mapping requirements
\citep{Kangas2018}. Indeed, the classical NFI can be supplemented by
AI-powered and all remote sensing-based products aforementioned,
through the concept of multi-Source forest inventory, an innovation
rewarded by the Marcus Wallenberg Prize in 1997 \citep{Tomppo2008} and
a matter of ongoing developments in France \citep{Vega2021,Sagar2022}.
A strong advantage here is that RS products play an auxiliary role in a
design-based statistical approach, and do not need to be unbiased, only
informative (i.e.\ correlated to the attribute of interest). 

Therefore, a virtuous loop can be figured out between NFI and RS
product developments, the former being used for calibrating the latter
in a high frequency and high-resolution perspective, while the latter
will usefully feed increased precision and frequency of the former
\citep{Bontemps2022}. Beyond the model-assisted paradigm, the
capability of AI to improve the correlation between remote sensing and
NFI will suppress the risk of bias associated with model specifications
\citep{Sagar2022} and enhance the capabilities of design-based and
model-assisted small-area estimation for estimating large disturbances.
For large scale (i.e., small area) disturbances, model-based small
area estimates remain an option \citep{Sagar2025} if no additional
field plot can be inventoried. In this specific field, research is also
ongoing to envision new and more adapted sampling strategies, using RS
products as a source of prior spatialized information on disturbances,
sampling strategies being directed by the former. A subsequent issue is
how to best articulate such approaches with standard forest inventory,
and trade the global sampling effort \citep{Gruijter2006}, in the
context of their expected increased occurrence \citep{Seidl2017}. The
approach may also serve the updating of inventory estimates after
disturbances.\looseness=1

Given the demonstrated complementarity of these AI-powered and remote
sensing-based models \citep{Besic2025,Sagar2022}, we can thus
confidently state that holistic monitoring approaches---combining
satellite observations, AI, and structured field data---will be
essential for improving mortality predictions and understanding forest
carbon dynamics and, for improving forest monitoring in general
\citep{Bontemps2022}.

\subsection{Paths for NFI-driven assessment of the long-term C forest
sink}\label{sec55}

The present aim of tracking changes in the forest carbon sink and its
components over a long-term period (1990--2022) has not been frequently
addressed to date, and it highlighted clear limitations associated with
the different tools contributing to this monitoring, for it to be
accurate and homogeneous over time. These limitations call for some
minimum caution in the interpretations of the results, and also come
along with additional recommendations to the community: 

(i) \textbf{Forest harvest reconstitutions over the long term}---before
2010, no accurate field monitoring of forest removals was implemented,
making the NFI data non-operative in this respect, and requiring
external and indirect quantification approaches (including the
\textit{Enqu\^{e}te Annuelle de Branche}, for wood sawn in the
industries) routinely using for forest carbon reporting. With forest
growing stock being however estimated with high precision (around 1.5\%
at country scale, and at a much higher precision in former decennial
inventories), it is suggested that forest harvests be retrospectively
reconstituted by equating the annual difference of carbon stocks (net
carbon sink) to the sum of forest fluxes. The approach has been
successfully tested over ancient departmental inventories 
\citep{DenardouTisserand2019}, and can be generalized to all available
inventories.

(ii)~\textbf{Delivering long time series of forest attributes}---Old
inventories were based on asynchronous departmental surveys, which form
a \mbox{systematic} limit for a rapid production of forest time series at
different scales at which statistical precision is sufficient
(GRECO/SER ecological classifications, or NUTS-3/NUTS-2 levels of
EU administrative units). Using interpolation and aggregations, time
series of forest state variables (area, total stock and biomass) and
fluxes (growth, harvest, mortality) should also be produced for these
different classifications. The linearity of these reconstitutions makes
the delivery of associated confidence intervals prone to rapid
development.

(iii)~\textbf{Updating the wood density component of the aerial forest carbon
sink}---Highly diversified forests of the French territory pose a sharp
limit on the accuracy of wood density quantification across tree
species (${>}150$ found at the censusable stage in France). While the
Carbofor project remains a cornerstone support, some density records
however remaining doubtful. Yet, an unprecedented effort relying on a
collaboration between IGN and INRAE has been consented since 2015, with
a view to establish a systematic wood density record for French
forests, based on the sampling design of the national forest inventory.
These open data \citep{Cuny2025} should be used whenever aerial tree
carbon stocks/sinks are requested, and may lead to a positive
reevaluation of forest carbon. Wood density models are now under
development that will capture the intraspecific diversity associated to
tree and stand attributes.

(iv) \textbf{Tracking for total natural tree mortality in forests}---Whereas
mortality was the matter of restricted concern for up to one
decade, its abrupt increase after severe drought or pest-related
disturbances has drawn attention onto this flux 
\citep[e.g.][]{Taccoen2019}. In spite of major improvements introduced in the
French NFI with the semi-permanent design, 5 years can remain limiting
for mortality assessment after massive events, with the risk of both
underestimating natural mortality, and overestimating planned harvests.
Complementary perspectives rely on both progress in forest inventory
design (e.g.\ spatially systematic panelling of plot revisits across 
5~years has no cost difference), and the development of density-dependent
mortality models for better separating demographic and environmental
mortality, and developments in remote sensing technologies, from which
one may expect bias-correction options in view of the higher frequency
of these products. For these reasons, it is urged to amplify the
scientific \mbox{dialog} between remote sensing and forest inventory science
on this topic.

\section{Concluding remarks and recommendations}\label{sec6}

We have shown that  French forests experienced an alarming decline of
their carbon sink which went from 49.3 MtCO$_{2}{\cdot}$yr$^{-1}$ in 1990 to
37.8 MtonCO$_{2}{\cdot}$yr$^{-1}$ in 2022. This decrease took place mainly
from 2013 to 2017 and seems to be associated with an abrupt rise of
both natural mortality from climate change and to a parallel rise of
harvests. This harvest increase could reflect salvaged trees after
natural dieback. In recent years, two northern regions have become net
carbon emitters to the atmosphere and other regions have carbon fluxes
that have decreased and are now close to zero. In contrast, southern
regions remained stable carbon sinks both in intensively managed
Atlantic forests in Les Landes and lightly managed Mediterranean
forests. These findings make the country's carbon neutrality target out
of reach for the current period as it was based in the initial 
SNBC (2020) from data from 2018 which defined as a target a sink of 41
MtCO$_{2}{\cdot}$yr$^{-1}$ in the forest sector in 2015, then revised to 39
MtCO$_{2}{\cdot}$yr$^{-1}$ in the second SNBC-{2} budget for 2019--2023. Even
this less ambitious objective became unachievable in light of the
continued decrease of the carbon sink, and has been revised in SNBC-3
down to 9 MtCO$_{2}{\cdot}$yr$^{-1}$ (Figure~\ref{fig1}c). The new objective
may still be challenging to reach in light of continuing greenhouse gas
emissions in the agriculture sub-sector and rising tree mortality
decreasing the forest sink in the coming years. 

Our findings show that tree mortality has become a key driver of the
declining carbon sink in France \citep[see also][]{HCC2025}, with
significant regional differences and species-specific trends. The
acceleration of mortality since 2013, independent of harvest trends,
suggesting that climate-induced stressors---especially droughts and
biotic disturbances---are reshaping forest dynamics. Given the delayed
detection of mortality in panelized annual national forest inventories,
complementing field-based observations with high-resolution remote
sensing and citizen science initiatives could provide earlier warning
signals and improve attribution of mortality drivers. In this context,
we strongly recommend that deadwood carbon stock changes caused by
increased mortality should be included in carbon accounting and
subsequent reports.

The national inventory is a unique resource to monitor long-term carbon
trends in forests and forms one of the bases for the France national
reporting to the UNFCCC. However, both the current sampling scheme and
the global sampling effort do not allow for year-to-year monitoring of
changes, and cannot attribute the epicenters and severity of carbon
losses occurring through disturbances as soon as these are small-scale.
Satellite imagery provides excellent coverage down to tree level but
requires ground data to provide robust maps of carbon accounting. We
recommend here to develop artificial intelligence and data integration
methods to monitor gains of forest carbon from satellite and airborne
imagery.

In this respect, the recent inclusion of the national forest inventory
in the framework of public statistics (under the authority of the CNIS
council) has led to specify and enforce data diffusion rules and
methods to both satisfy the needs of public research and the
statistical secrecy law, permitting in particular the exact spatial
matching of high-resolution satellite products and field data.
Additional recommendations can be formulated, including the provision
of geospatial information on forest clearcuts, and natural disturbances
such as insects, drought mortality events based on the best available
science, the measure of deadwood changes as an overlooked carbon stock
in French forests, the development of artificial intelligence and data
integration methods to monitor gains or losses of forest carbon from
satellite and airborne imagery.

\section*{Acknowledgements}
PC acknowledges support from the EYE CLIMA HE project
Grant~Agreement No. 101081395, the French German AI4FOREST project
(ANR-22-FAI1-0002-01). PC, NB and CV from the CNES
TOSCA FOREST-C50m project.

\CDRGrant[EYE CLIMA HE]{101081395}
\CDRGrant[ANR]{ANR-22-FAI1-0002-01}

\section*{Declaration of interests}
The authors do not work for, advise, own shares in, or receive funds
from any organization that could benefit from this article, and have
declared no affiliations other than their research organizations.

\back{}

\printbibliography
\refinput{crgeos20250220-reference.tex}

\end{document}